yading@10: /* yading@10: * AMR wideband decoder yading@10: * Copyright (c) 2010 Marcelo Galvao Povoa yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A particular PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * AMR wideband decoder yading@10: */ yading@10: yading@10: #include "libavutil/channel_layout.h" yading@10: #include "libavutil/common.h" yading@10: #include "libavutil/float_dsp.h" yading@10: #include "libavutil/lfg.h" yading@10: yading@10: #include "avcodec.h" yading@10: #include "lsp.h" yading@10: #include "celp_filters.h" yading@10: #include "celp_math.h" yading@10: #include "acelp_filters.h" yading@10: #include "acelp_vectors.h" yading@10: #include "acelp_pitch_delay.h" yading@10: #include "internal.h" yading@10: yading@10: #define AMR_USE_16BIT_TABLES yading@10: #include "amr.h" yading@10: yading@10: #include "amrwbdata.h" yading@10: #include "mips/amrwbdec_mips.h" yading@10: yading@10: typedef struct { yading@10: AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream yading@10: enum Mode fr_cur_mode; ///< mode index of current frame yading@10: uint8_t fr_quality; ///< frame quality index (FQI) yading@10: float isf_cur[LP_ORDER]; ///< working ISF vector from current frame yading@10: float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame yading@10: float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame yading@10: double isp[4][LP_ORDER]; ///< ISP vectors from current frame yading@10: double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame yading@10: yading@10: float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector yading@10: yading@10: uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe yading@10: uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe yading@10: yading@10: float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history yading@10: float *excitation; ///< points to current excitation in excitation_buf[] yading@10: yading@10: float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe yading@10: float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe yading@10: yading@10: float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes yading@10: float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes yading@10: float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes yading@10: yading@10: float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe yading@10: yading@10: float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset" yading@10: uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none yading@10: float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold yading@10: yading@10: float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz yading@10: float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling yading@10: float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz yading@10: yading@10: float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters yading@10: float demph_mem[1]; ///< previous value in the de-emphasis filter yading@10: float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter yading@10: float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter yading@10: yading@10: AVLFG prng; ///< random number generator for white noise excitation yading@10: uint8_t first_frame; ///< flag active during decoding of the first frame yading@10: ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs yading@10: ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs yading@10: CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs yading@10: CELPMContext celpm_ctx; ///< context for fixed point math operations yading@10: yading@10: } AMRWBContext; yading@10: yading@10: static av_cold int amrwb_decode_init(AVCodecContext *avctx) yading@10: { yading@10: AMRWBContext *ctx = avctx->priv_data; yading@10: int i; yading@10: yading@10: if (avctx->channels > 1) { yading@10: avpriv_report_missing_feature(avctx, "multi-channel AMR"); yading@10: return AVERROR_PATCHWELCOME; yading@10: } yading@10: yading@10: avctx->channels = 1; yading@10: avctx->channel_layout = AV_CH_LAYOUT_MONO; yading@10: if (!avctx->sample_rate) yading@10: avctx->sample_rate = 16000; yading@10: avctx->sample_fmt = AV_SAMPLE_FMT_FLT; yading@10: yading@10: av_lfg_init(&ctx->prng, 1); yading@10: yading@10: ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1]; yading@10: ctx->first_frame = 1; yading@10: yading@10: for (i = 0; i < LP_ORDER; i++) yading@10: ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: ctx->prediction_error[i] = MIN_ENERGY; yading@10: yading@10: ff_acelp_filter_init(&ctx->acelpf_ctx); yading@10: ff_acelp_vectors_init(&ctx->acelpv_ctx); yading@10: ff_celp_filter_init(&ctx->celpf_ctx); yading@10: ff_celp_math_init(&ctx->celpm_ctx); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Decode the frame header in the "MIME/storage" format. This format yading@10: * is simpler and does not carry the auxiliary frame information. yading@10: * yading@10: * @param[in] ctx The Context yading@10: * @param[in] buf Pointer to the input buffer yading@10: * yading@10: * @return The decoded header length in bytes yading@10: */ yading@10: static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf) yading@10: { yading@10: /* Decode frame header (1st octet) */ yading@10: ctx->fr_cur_mode = buf[0] >> 3 & 0x0F; yading@10: ctx->fr_quality = (buf[0] & 0x4) == 0x4; yading@10: yading@10: return 1; yading@10: } yading@10: yading@10: /** yading@10: * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only). yading@10: * yading@10: * @param[in] ind Array of 5 indexes yading@10: * @param[out] isf_q Buffer for isf_q[LP_ORDER] yading@10: * yading@10: */ yading@10: static void decode_isf_indices_36b(uint16_t *ind, float *isf_q) yading@10: { yading@10: int i; yading@10: yading@10: for (i = 0; i < 9; i++) yading@10: isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 7; i++) yading@10: isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 5; i++) yading@10: isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 7; i++) yading@10: isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15)); yading@10: } yading@10: yading@10: /** yading@10: * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode). yading@10: * yading@10: * @param[in] ind Array of 7 indexes yading@10: * @param[out] isf_q Buffer for isf_q[LP_ORDER] yading@10: * yading@10: */ yading@10: static void decode_isf_indices_46b(uint16_t *ind, float *isf_q) yading@10: { yading@10: int i; yading@10: yading@10: for (i = 0; i < 9; i++) yading@10: isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 7; i++) yading@10: isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 3; i++) yading@10: isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 3; i++) yading@10: isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 3; i++) yading@10: isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 3; i++) yading@10: isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15)); yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15)); yading@10: } yading@10: yading@10: /** yading@10: * Apply mean and past ISF values using the prediction factor. yading@10: * Updates past ISF vector. yading@10: * yading@10: * @param[in,out] isf_q Current quantized ISF yading@10: * @param[in,out] isf_past Past quantized ISF yading@10: * yading@10: */ yading@10: static void isf_add_mean_and_past(float *isf_q, float *isf_past) yading@10: { yading@10: int i; yading@10: float tmp; yading@10: yading@10: for (i = 0; i < LP_ORDER; i++) { yading@10: tmp = isf_q[i]; yading@10: isf_q[i] += isf_mean[i] * (1.0f / (1 << 15)); yading@10: isf_q[i] += PRED_FACTOR * isf_past[i]; yading@10: isf_past[i] = tmp; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Interpolate the fourth ISP vector from current and past frames yading@10: * to obtain an ISP vector for each subframe. yading@10: * yading@10: * @param[in,out] isp_q ISPs for each subframe yading@10: * @param[in] isp4_past Past ISP for subframe 4 yading@10: */ yading@10: static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past) yading@10: { yading@10: int i, k; yading@10: yading@10: for (k = 0; k < 3; k++) { yading@10: float c = isfp_inter[k]; yading@10: for (i = 0; i < LP_ORDER; i++) yading@10: isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i]; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes). yading@10: * Calculate integer lag and fractional lag always using 1/4 resolution. yading@10: * In 1st and 3rd subframes the index is relative to last subframe integer lag. yading@10: * yading@10: * @param[out] lag_int Decoded integer pitch lag yading@10: * @param[out] lag_frac Decoded fractional pitch lag yading@10: * @param[in] pitch_index Adaptive codebook pitch index yading@10: * @param[in,out] base_lag_int Base integer lag used in relative subframes yading@10: * @param[in] subframe Current subframe index (0 to 3) yading@10: */ yading@10: static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index, yading@10: uint8_t *base_lag_int, int subframe) yading@10: { yading@10: if (subframe == 0 || subframe == 2) { yading@10: if (pitch_index < 376) { yading@10: *lag_int = (pitch_index + 137) >> 2; yading@10: *lag_frac = pitch_index - (*lag_int << 2) + 136; yading@10: } else if (pitch_index < 440) { yading@10: *lag_int = (pitch_index + 257 - 376) >> 1; yading@10: *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1; yading@10: /* the actual resolution is 1/2 but expressed as 1/4 */ yading@10: } else { yading@10: *lag_int = pitch_index - 280; yading@10: *lag_frac = 0; yading@10: } yading@10: /* minimum lag for next subframe */ yading@10: *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0), yading@10: AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15); yading@10: // XXX: the spec states clearly that *base_lag_int should be yading@10: // the nearest integer to *lag_int (minus 8), but the ref code yading@10: // actually always uses its floor, I'm following the latter yading@10: } else { yading@10: *lag_int = (pitch_index + 1) >> 2; yading@10: *lag_frac = pitch_index - (*lag_int << 2); yading@10: *lag_int += *base_lag_int; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes. yading@10: * The description is analogous to decode_pitch_lag_high, but in 6k60 the yading@10: * relative index is used for all subframes except the first. yading@10: */ yading@10: static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index, yading@10: uint8_t *base_lag_int, int subframe, enum Mode mode) yading@10: { yading@10: if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) { yading@10: if (pitch_index < 116) { yading@10: *lag_int = (pitch_index + 69) >> 1; yading@10: *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1; yading@10: } else { yading@10: *lag_int = pitch_index - 24; yading@10: *lag_frac = 0; yading@10: } yading@10: // XXX: same problem as before yading@10: *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0), yading@10: AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15); yading@10: } else { yading@10: *lag_int = (pitch_index + 1) >> 1; yading@10: *lag_frac = (pitch_index - (*lag_int << 1)) << 1; yading@10: *lag_int += *base_lag_int; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Find the pitch vector by interpolating the past excitation at the yading@10: * pitch delay, which is obtained in this function. yading@10: * yading@10: * @param[in,out] ctx The context yading@10: * @param[in] amr_subframe Current subframe data yading@10: * @param[in] subframe Current subframe index (0 to 3) yading@10: */ yading@10: static void decode_pitch_vector(AMRWBContext *ctx, yading@10: const AMRWBSubFrame *amr_subframe, yading@10: const int subframe) yading@10: { yading@10: int pitch_lag_int, pitch_lag_frac; yading@10: int i; yading@10: float *exc = ctx->excitation; yading@10: enum Mode mode = ctx->fr_cur_mode; yading@10: yading@10: if (mode <= MODE_8k85) { yading@10: decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap, yading@10: &ctx->base_pitch_lag, subframe, mode); yading@10: } else yading@10: decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap, yading@10: &ctx->base_pitch_lag, subframe); yading@10: yading@10: ctx->pitch_lag_int = pitch_lag_int; yading@10: pitch_lag_int += pitch_lag_frac > 0; yading@10: yading@10: /* Calculate the pitch vector by interpolating the past excitation at the yading@10: pitch lag using a hamming windowed sinc function */ yading@10: ctx->acelpf_ctx.acelp_interpolatef(exc, yading@10: exc + 1 - pitch_lag_int, yading@10: ac_inter, 4, yading@10: pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4), yading@10: LP_ORDER, AMRWB_SFR_SIZE + 1); yading@10: yading@10: /* Check which pitch signal path should be used yading@10: * 6k60 and 8k85 modes have the ltp flag set to 0 */ yading@10: if (amr_subframe->ltp) { yading@10: memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float)); yading@10: } else { yading@10: for (i = 0; i < AMRWB_SFR_SIZE; i++) yading@10: ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] + yading@10: 0.18 * exc[i + 1]; yading@10: memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float)); yading@10: } yading@10: } yading@10: yading@10: /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */ yading@10: #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1)) yading@10: yading@10: /** Get the bit at specified position */ yading@10: #define BIT_POS(x, p) (((x) >> (p)) & 1) yading@10: yading@10: /** yading@10: * The next six functions decode_[i]p_track decode exactly i pulses yading@10: * positions and amplitudes (-1 or 1) in a subframe track using yading@10: * an encoded pulse indexing (TS 26.190 section 5.8.2). yading@10: * yading@10: * The results are given in out[], in which a negative number means yading@10: * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ). yading@10: * yading@10: * @param[out] out Output buffer (writes i elements) yading@10: * @param[in] code Pulse index (no. of bits varies, see below) yading@10: * @param[in] m (log2) Number of potential positions yading@10: * @param[in] off Offset for decoded positions yading@10: */ yading@10: static inline void decode_1p_track(int *out, int code, int m, int off) yading@10: { yading@10: int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits yading@10: yading@10: out[0] = BIT_POS(code, m) ? -pos : pos; yading@10: } yading@10: yading@10: static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits yading@10: { yading@10: int pos0 = BIT_STR(code, m, m) + off; yading@10: int pos1 = BIT_STR(code, 0, m) + off; yading@10: yading@10: out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0; yading@10: out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1; yading@10: out[1] = pos0 > pos1 ? -out[1] : out[1]; yading@10: } yading@10: yading@10: static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits yading@10: { yading@10: int half_2p = BIT_POS(code, 2*m - 1) << (m - 1); yading@10: yading@10: decode_2p_track(out, BIT_STR(code, 0, 2*m - 1), yading@10: m - 1, off + half_2p); yading@10: decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off); yading@10: } yading@10: yading@10: static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits yading@10: { yading@10: int half_4p, subhalf_2p; yading@10: int b_offset = 1 << (m - 1); yading@10: yading@10: switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */ yading@10: case 0: /* 0 pulses in A, 4 pulses in B or vice versa */ yading@10: half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses yading@10: subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2); yading@10: yading@10: decode_2p_track(out, BIT_STR(code, 0, 2*m - 3), yading@10: m - 2, off + half_4p + subhalf_2p); yading@10: decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1), yading@10: m - 1, off + half_4p); yading@10: break; yading@10: case 1: /* 1 pulse in A, 3 pulses in B */ yading@10: decode_1p_track(out, BIT_STR(code, 3*m - 2, m), yading@10: m - 1, off); yading@10: decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2), yading@10: m - 1, off + b_offset); yading@10: break; yading@10: case 2: /* 2 pulses in each half */ yading@10: decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1), yading@10: m - 1, off); yading@10: decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1), yading@10: m - 1, off + b_offset); yading@10: break; yading@10: case 3: /* 3 pulses in A, 1 pulse in B */ yading@10: decode_3p_track(out, BIT_STR(code, m, 3*m - 2), yading@10: m - 1, off); yading@10: decode_1p_track(out + 3, BIT_STR(code, 0, m), yading@10: m - 1, off + b_offset); yading@10: break; yading@10: } yading@10: } yading@10: yading@10: static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits yading@10: { yading@10: int half_3p = BIT_POS(code, 5*m - 1) << (m - 1); yading@10: yading@10: decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2), yading@10: m - 1, off + half_3p); yading@10: yading@10: decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off); yading@10: } yading@10: yading@10: static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits yading@10: { yading@10: int b_offset = 1 << (m - 1); yading@10: /* which half has more pulses in cases 0 to 2 */ yading@10: int half_more = BIT_POS(code, 6*m - 5) << (m - 1); yading@10: int half_other = b_offset - half_more; yading@10: yading@10: switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */ yading@10: case 0: /* 0 pulses in A, 6 pulses in B or vice versa */ yading@10: decode_1p_track(out, BIT_STR(code, 0, m), yading@10: m - 1, off + half_more); yading@10: decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5), yading@10: m - 1, off + half_more); yading@10: break; yading@10: case 1: /* 1 pulse in A, 5 pulses in B or vice versa */ yading@10: decode_1p_track(out, BIT_STR(code, 0, m), yading@10: m - 1, off + half_other); yading@10: decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5), yading@10: m - 1, off + half_more); yading@10: break; yading@10: case 2: /* 2 pulses in A, 4 pulses in B or vice versa */ yading@10: decode_2p_track(out, BIT_STR(code, 0, 2*m - 1), yading@10: m - 1, off + half_other); yading@10: decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4), yading@10: m - 1, off + half_more); yading@10: break; yading@10: case 3: /* 3 pulses in A, 3 pulses in B */ yading@10: decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2), yading@10: m - 1, off); yading@10: decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2), yading@10: m - 1, off + b_offset); yading@10: break; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Decode the algebraic codebook index to pulse positions and signs, yading@10: * then construct the algebraic codebook vector. yading@10: * yading@10: * @param[out] fixed_vector Buffer for the fixed codebook excitation yading@10: * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only) yading@10: * @param[in] pulse_lo LSBs part of the pulse index array yading@10: * @param[in] mode Mode of the current frame yading@10: */ yading@10: static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi, yading@10: const uint16_t *pulse_lo, const enum Mode mode) yading@10: { yading@10: /* sig_pos stores for each track the decoded pulse position indexes yading@10: * (1-based) multiplied by its corresponding amplitude (+1 or -1) */ yading@10: int sig_pos[4][6]; yading@10: int spacing = (mode == MODE_6k60) ? 2 : 4; yading@10: int i, j; yading@10: yading@10: switch (mode) { yading@10: case MODE_6k60: yading@10: for (i = 0; i < 2; i++) yading@10: decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1); yading@10: break; yading@10: case MODE_8k85: yading@10: for (i = 0; i < 4; i++) yading@10: decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1); yading@10: break; yading@10: case MODE_12k65: yading@10: for (i = 0; i < 4; i++) yading@10: decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1); yading@10: break; yading@10: case MODE_14k25: yading@10: for (i = 0; i < 2; i++) yading@10: decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1); yading@10: for (i = 2; i < 4; i++) yading@10: decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1); yading@10: break; yading@10: case MODE_15k85: yading@10: for (i = 0; i < 4; i++) yading@10: decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1); yading@10: break; yading@10: case MODE_18k25: yading@10: for (i = 0; i < 4; i++) yading@10: decode_4p_track(sig_pos[i], (int) pulse_lo[i] + yading@10: ((int) pulse_hi[i] << 14), 4, 1); yading@10: break; yading@10: case MODE_19k85: yading@10: for (i = 0; i < 2; i++) yading@10: decode_5p_track(sig_pos[i], (int) pulse_lo[i] + yading@10: ((int) pulse_hi[i] << 10), 4, 1); yading@10: for (i = 2; i < 4; i++) yading@10: decode_4p_track(sig_pos[i], (int) pulse_lo[i] + yading@10: ((int) pulse_hi[i] << 14), 4, 1); yading@10: break; yading@10: case MODE_23k05: yading@10: case MODE_23k85: yading@10: for (i = 0; i < 4; i++) yading@10: decode_6p_track(sig_pos[i], (int) pulse_lo[i] + yading@10: ((int) pulse_hi[i] << 11), 4, 1); yading@10: break; yading@10: } yading@10: yading@10: memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE); yading@10: yading@10: for (i = 0; i < 4; i++) yading@10: for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) { yading@10: int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i; yading@10: yading@10: fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Decode pitch gain and fixed gain correction factor. yading@10: * yading@10: * @param[in] vq_gain Vector-quantized index for gains yading@10: * @param[in] mode Mode of the current frame yading@10: * @param[out] fixed_gain_factor Decoded fixed gain correction factor yading@10: * @param[out] pitch_gain Decoded pitch gain yading@10: */ yading@10: static void decode_gains(const uint8_t vq_gain, const enum Mode mode, yading@10: float *fixed_gain_factor, float *pitch_gain) yading@10: { yading@10: const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] : yading@10: qua_gain_7b[vq_gain]); yading@10: yading@10: *pitch_gain = gains[0] * (1.0f / (1 << 14)); yading@10: *fixed_gain_factor = gains[1] * (1.0f / (1 << 11)); yading@10: } yading@10: yading@10: /** yading@10: * Apply pitch sharpening filters to the fixed codebook vector. yading@10: * yading@10: * @param[in] ctx The context yading@10: * @param[in,out] fixed_vector Fixed codebook excitation yading@10: */ yading@10: // XXX: Spec states this procedure should be applied when the pitch yading@10: // lag is less than 64, but this checking seems absent in reference and AMR-NB yading@10: static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector) yading@10: { yading@10: int i; yading@10: yading@10: /* Tilt part */ yading@10: for (i = AMRWB_SFR_SIZE - 1; i != 0; i--) yading@10: fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef; yading@10: yading@10: /* Periodicity enhancement part */ yading@10: for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++) yading@10: fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85; yading@10: } yading@10: yading@10: /** yading@10: * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced). yading@10: * yading@10: * @param[in] p_vector, f_vector Pitch and fixed excitation vectors yading@10: * @param[in] p_gain, f_gain Pitch and fixed gains yading@10: * @param[in] ctx The context yading@10: */ yading@10: // XXX: There is something wrong with the precision here! The magnitudes yading@10: // of the energies are not correct. Please check the reference code carefully yading@10: static float voice_factor(float *p_vector, float p_gain, yading@10: float *f_vector, float f_gain, yading@10: CELPMContext *ctx) yading@10: { yading@10: double p_ener = (double) ctx->dot_productf(p_vector, p_vector, yading@10: AMRWB_SFR_SIZE) * yading@10: p_gain * p_gain; yading@10: double f_ener = (double) ctx->dot_productf(f_vector, f_vector, yading@10: AMRWB_SFR_SIZE) * yading@10: f_gain * f_gain; yading@10: yading@10: return (p_ener - f_ener) / (p_ener + f_ener); yading@10: } yading@10: yading@10: /** yading@10: * Reduce fixed vector sparseness by smoothing with one of three IR filters, yading@10: * also known as "adaptive phase dispersion". yading@10: * yading@10: * @param[in] ctx The context yading@10: * @param[in,out] fixed_vector Unfiltered fixed vector yading@10: * @param[out] buf Space for modified vector if necessary yading@10: * yading@10: * @return The potentially overwritten filtered fixed vector address yading@10: */ yading@10: static float *anti_sparseness(AMRWBContext *ctx, yading@10: float *fixed_vector, float *buf) yading@10: { yading@10: int ir_filter_nr; yading@10: yading@10: if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes yading@10: return fixed_vector; yading@10: yading@10: if (ctx->pitch_gain[0] < 0.6) { yading@10: ir_filter_nr = 0; // strong filtering yading@10: } else if (ctx->pitch_gain[0] < 0.9) { yading@10: ir_filter_nr = 1; // medium filtering yading@10: } else yading@10: ir_filter_nr = 2; // no filtering yading@10: yading@10: /* detect 'onset' */ yading@10: if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) { yading@10: if (ir_filter_nr < 2) yading@10: ir_filter_nr++; yading@10: } else { yading@10: int i, count = 0; yading@10: yading@10: for (i = 0; i < 6; i++) yading@10: if (ctx->pitch_gain[i] < 0.6) yading@10: count++; yading@10: yading@10: if (count > 2) yading@10: ir_filter_nr = 0; yading@10: yading@10: if (ir_filter_nr > ctx->prev_ir_filter_nr + 1) yading@10: ir_filter_nr--; yading@10: } yading@10: yading@10: /* update ir filter strength history */ yading@10: ctx->prev_ir_filter_nr = ir_filter_nr; yading@10: yading@10: ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85); yading@10: yading@10: if (ir_filter_nr < 2) { yading@10: int i; yading@10: const float *coef = ir_filters_lookup[ir_filter_nr]; yading@10: yading@10: /* Circular convolution code in the reference yading@10: * decoder was modified to avoid using one yading@10: * extra array. The filtered vector is given by: yading@10: * yading@10: * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) } yading@10: */ yading@10: yading@10: memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE); yading@10: for (i = 0; i < AMRWB_SFR_SIZE; i++) yading@10: if (fixed_vector[i]) yading@10: ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i], yading@10: AMRWB_SFR_SIZE); yading@10: fixed_vector = buf; yading@10: } yading@10: yading@10: return fixed_vector; yading@10: } yading@10: yading@10: /** yading@10: * Calculate a stability factor {teta} based on distance between yading@10: * current and past isf. A value of 1 shows maximum signal stability. yading@10: */ yading@10: static float stability_factor(const float *isf, const float *isf_past) yading@10: { yading@10: int i; yading@10: float acc = 0.0; yading@10: yading@10: for (i = 0; i < LP_ORDER - 1; i++) yading@10: acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]); yading@10: yading@10: // XXX: This part is not so clear from the reference code yading@10: // the result is more accurate changing the "/ 256" to "* 512" yading@10: return FFMAX(0.0, 1.25 - acc * 0.8 * 512); yading@10: } yading@10: yading@10: /** yading@10: * Apply a non-linear fixed gain smoothing in order to reduce yading@10: * fluctuation in the energy of excitation. yading@10: * yading@10: * @param[in] fixed_gain Unsmoothed fixed gain yading@10: * @param[in,out] prev_tr_gain Previous threshold gain (updated) yading@10: * @param[in] voice_fac Frame voicing factor yading@10: * @param[in] stab_fac Frame stability factor yading@10: * yading@10: * @return The smoothed gain yading@10: */ yading@10: static float noise_enhancer(float fixed_gain, float *prev_tr_gain, yading@10: float voice_fac, float stab_fac) yading@10: { yading@10: float sm_fac = 0.5 * (1 - voice_fac) * stab_fac; yading@10: float g0; yading@10: yading@10: // XXX: the following fixed-point constants used to in(de)crement yading@10: // gain by 1.5dB were taken from the reference code, maybe it could yading@10: // be simpler yading@10: if (fixed_gain < *prev_tr_gain) { yading@10: g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain * yading@10: (6226 * (1.0f / (1 << 15)))); // +1.5 dB yading@10: } else yading@10: g0 = FFMAX(*prev_tr_gain, fixed_gain * yading@10: (27536 * (1.0f / (1 << 15)))); // -1.5 dB yading@10: yading@10: *prev_tr_gain = g0; // update next frame threshold yading@10: yading@10: return sm_fac * g0 + (1 - sm_fac) * fixed_gain; yading@10: } yading@10: yading@10: /** yading@10: * Filter the fixed_vector to emphasize the higher frequencies. yading@10: * yading@10: * @param[in,out] fixed_vector Fixed codebook vector yading@10: * @param[in] voice_fac Frame voicing factor yading@10: */ yading@10: static void pitch_enhancer(float *fixed_vector, float voice_fac) yading@10: { yading@10: int i; yading@10: float cpe = 0.125 * (1 + voice_fac); yading@10: float last = fixed_vector[0]; // holds c(i - 1) yading@10: yading@10: fixed_vector[0] -= cpe * fixed_vector[1]; yading@10: yading@10: for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) { yading@10: float cur = fixed_vector[i]; yading@10: yading@10: fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]); yading@10: last = cur; yading@10: } yading@10: yading@10: fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last; yading@10: } yading@10: yading@10: /** yading@10: * Conduct 16th order linear predictive coding synthesis from excitation. yading@10: * yading@10: * @param[in] ctx Pointer to the AMRWBContext yading@10: * @param[in] lpc Pointer to the LPC coefficients yading@10: * @param[out] excitation Buffer for synthesis final excitation yading@10: * @param[in] fixed_gain Fixed codebook gain for synthesis yading@10: * @param[in] fixed_vector Algebraic codebook vector yading@10: * @param[in,out] samples Pointer to the output samples and memory yading@10: */ yading@10: static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation, yading@10: float fixed_gain, const float *fixed_vector, yading@10: float *samples) yading@10: { yading@10: ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector, yading@10: ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE); yading@10: yading@10: /* emphasize pitch vector contribution in low bitrate modes */ yading@10: if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) { yading@10: int i; yading@10: float energy = ctx->celpm_ctx.dot_productf(excitation, excitation, yading@10: AMRWB_SFR_SIZE); yading@10: yading@10: // XXX: Weird part in both ref code and spec. A unknown parameter yading@10: // {beta} seems to be identical to the current pitch gain yading@10: float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0]; yading@10: yading@10: for (i = 0; i < AMRWB_SFR_SIZE; i++) yading@10: excitation[i] += pitch_factor * ctx->pitch_vector[i]; yading@10: yading@10: ff_scale_vector_to_given_sum_of_squares(excitation, excitation, yading@10: energy, AMRWB_SFR_SIZE); yading@10: } yading@10: yading@10: ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation, yading@10: AMRWB_SFR_SIZE, LP_ORDER); yading@10: } yading@10: yading@10: /** yading@10: * Apply to synthesis a de-emphasis filter of the form: yading@10: * H(z) = 1 / (1 - m * z^-1) yading@10: * yading@10: * @param[out] out Output buffer yading@10: * @param[in] in Input samples array with in[-1] yading@10: * @param[in] m Filter coefficient yading@10: * @param[in,out] mem State from last filtering yading@10: */ yading@10: static void de_emphasis(float *out, float *in, float m, float mem[1]) yading@10: { yading@10: int i; yading@10: yading@10: out[0] = in[0] + m * mem[0]; yading@10: yading@10: for (i = 1; i < AMRWB_SFR_SIZE; i++) yading@10: out[i] = in[i] + out[i - 1] * m; yading@10: yading@10: mem[0] = out[AMRWB_SFR_SIZE - 1]; yading@10: } yading@10: yading@10: /** yading@10: * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using yading@10: * a FIR interpolation filter. Uses past data from before *in address. yading@10: * yading@10: * @param[out] out Buffer for interpolated signal yading@10: * @param[in] in Current signal data (length 0.8*o_size) yading@10: * @param[in] o_size Output signal length yading@10: * @param[in] ctx The context yading@10: */ yading@10: static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx) yading@10: { yading@10: const float *in0 = in - UPS_FIR_SIZE + 1; yading@10: int i, j, k; yading@10: int int_part = 0, frac_part; yading@10: yading@10: i = 0; yading@10: for (j = 0; j < o_size / 5; j++) { yading@10: out[i] = in[int_part]; yading@10: frac_part = 4; yading@10: i++; yading@10: yading@10: for (k = 1; k < 5; k++) { yading@10: out[i] = ctx->dot_productf(in0 + int_part, yading@10: upsample_fir[4 - frac_part], yading@10: UPS_MEM_SIZE); yading@10: int_part++; yading@10: frac_part--; yading@10: i++; yading@10: } yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Calculate the high-band gain based on encoded index (23k85 mode) or yading@10: * on the low-band speech signal and the Voice Activity Detection flag. yading@10: * yading@10: * @param[in] ctx The context yading@10: * @param[in] synth LB speech synthesis at 12.8k yading@10: * @param[in] hb_idx Gain index for mode 23k85 only yading@10: * @param[in] vad VAD flag for the frame yading@10: */ yading@10: static float find_hb_gain(AMRWBContext *ctx, const float *synth, yading@10: uint16_t hb_idx, uint8_t vad) yading@10: { yading@10: int wsp = (vad > 0); yading@10: float tilt; yading@10: yading@10: if (ctx->fr_cur_mode == MODE_23k85) yading@10: return qua_hb_gain[hb_idx] * (1.0f / (1 << 14)); yading@10: yading@10: tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) / yading@10: ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE); yading@10: yading@10: /* return gain bounded by [0.1, 1.0] */ yading@10: return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0); yading@10: } yading@10: yading@10: /** yading@10: * Generate the high-band excitation with the same energy from the lower yading@10: * one and scaled by the given gain. yading@10: * yading@10: * @param[in] ctx The context yading@10: * @param[out] hb_exc Buffer for the excitation yading@10: * @param[in] synth_exc Low-band excitation used for synthesis yading@10: * @param[in] hb_gain Wanted excitation gain yading@10: */ yading@10: static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc, yading@10: const float *synth_exc, float hb_gain) yading@10: { yading@10: int i; yading@10: float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, yading@10: AMRWB_SFR_SIZE); yading@10: yading@10: /* Generate a white-noise excitation */ yading@10: for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) yading@10: hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng); yading@10: yading@10: ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc, yading@10: energy * hb_gain * hb_gain, yading@10: AMRWB_SFR_SIZE_16k); yading@10: } yading@10: yading@10: /** yading@10: * Calculate the auto-correlation for the ISF difference vector. yading@10: */ yading@10: static float auto_correlation(float *diff_isf, float mean, int lag) yading@10: { yading@10: int i; yading@10: float sum = 0.0; yading@10: yading@10: for (i = 7; i < LP_ORDER - 2; i++) { yading@10: float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean); yading@10: sum += prod * prod; yading@10: } yading@10: return sum; yading@10: } yading@10: yading@10: /** yading@10: * Extrapolate a ISF vector to the 16kHz range (20th order LP) yading@10: * used at mode 6k60 LP filter for the high frequency band. yading@10: * yading@10: * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER yading@10: * values on input yading@10: */ yading@10: static void extrapolate_isf(float isf[LP_ORDER_16k]) yading@10: { yading@10: float diff_isf[LP_ORDER - 2], diff_mean; yading@10: float corr_lag[3]; yading@10: float est, scale; yading@10: int i, j, i_max_corr; yading@10: yading@10: isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1]; yading@10: yading@10: /* Calculate the difference vector */ yading@10: for (i = 0; i < LP_ORDER - 2; i++) yading@10: diff_isf[i] = isf[i + 1] - isf[i]; yading@10: yading@10: diff_mean = 0.0; yading@10: for (i = 2; i < LP_ORDER - 2; i++) yading@10: diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4)); yading@10: yading@10: /* Find which is the maximum autocorrelation */ yading@10: i_max_corr = 0; yading@10: for (i = 0; i < 3; i++) { yading@10: corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2); yading@10: yading@10: if (corr_lag[i] > corr_lag[i_max_corr]) yading@10: i_max_corr = i; yading@10: } yading@10: i_max_corr++; yading@10: yading@10: for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++) yading@10: isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr] yading@10: - isf[i - 2 - i_max_corr]; yading@10: yading@10: /* Calculate an estimate for ISF(18) and scale ISF based on the error */ yading@10: est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0; yading@10: scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) / yading@10: (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]); yading@10: yading@10: for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++) yading@10: diff_isf[j] = scale * (isf[i] - isf[i - 1]); yading@10: yading@10: /* Stability insurance */ yading@10: for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++) yading@10: if (diff_isf[i] + diff_isf[i - 1] < 5.0) { yading@10: if (diff_isf[i] > diff_isf[i - 1]) { yading@10: diff_isf[i - 1] = 5.0 - diff_isf[i]; yading@10: } else yading@10: diff_isf[i] = 5.0 - diff_isf[i - 1]; yading@10: } yading@10: yading@10: for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++) yading@10: isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15)); yading@10: yading@10: /* Scale the ISF vector for 16000 Hz */ yading@10: for (i = 0; i < LP_ORDER_16k - 1; i++) yading@10: isf[i] *= 0.8; yading@10: } yading@10: yading@10: /** yading@10: * Spectral expand the LP coefficients using the equation: yading@10: * y[i] = x[i] * (gamma ** i) yading@10: * yading@10: * @param[out] out Output buffer (may use input array) yading@10: * @param[in] lpc LP coefficients array yading@10: * @param[in] gamma Weighting factor yading@10: * @param[in] size LP array size yading@10: */ yading@10: static void lpc_weighting(float *out, const float *lpc, float gamma, int size) yading@10: { yading@10: int i; yading@10: float fac = gamma; yading@10: yading@10: for (i = 0; i < size; i++) { yading@10: out[i] = lpc[i] * fac; yading@10: fac *= gamma; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Conduct 20th order linear predictive coding synthesis for the high yading@10: * frequency band excitation at 16kHz. yading@10: * yading@10: * @param[in] ctx The context yading@10: * @param[in] subframe Current subframe index (0 to 3) yading@10: * @param[in,out] samples Pointer to the output speech samples yading@10: * @param[in] exc Generated white-noise scaled excitation yading@10: * @param[in] isf Current frame isf vector yading@10: * @param[in] isf_past Past frame final isf vector yading@10: */ yading@10: static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples, yading@10: const float *exc, const float *isf, const float *isf_past) yading@10: { yading@10: float hb_lpc[LP_ORDER_16k]; yading@10: enum Mode mode = ctx->fr_cur_mode; yading@10: yading@10: if (mode == MODE_6k60) { yading@10: float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation yading@10: double e_isp[LP_ORDER_16k]; yading@10: yading@10: ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe], yading@10: 1.0 - isfp_inter[subframe], LP_ORDER); yading@10: yading@10: extrapolate_isf(e_isf); yading@10: yading@10: e_isf[LP_ORDER_16k - 1] *= 2.0; yading@10: ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k); yading@10: ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k); yading@10: yading@10: lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k); yading@10: } else { yading@10: lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER); yading@10: } yading@10: yading@10: ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k, yading@10: (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER); yading@10: } yading@10: yading@10: /** yading@10: * Apply a 15th order filter to high-band samples. yading@10: * The filter characteristic depends on the given coefficients. yading@10: * yading@10: * @param[out] out Buffer for filtered output yading@10: * @param[in] fir_coef Filter coefficients yading@10: * @param[in,out] mem State from last filtering (updated) yading@10: * @param[in] in Input speech data (high-band) yading@10: * yading@10: * @remark It is safe to pass the same array in in and out parameters yading@10: */ yading@10: yading@10: #ifndef hb_fir_filter yading@10: static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1], yading@10: float mem[HB_FIR_SIZE], const float *in) yading@10: { yading@10: int i, j; yading@10: float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples yading@10: yading@10: memcpy(data, mem, HB_FIR_SIZE * sizeof(float)); yading@10: memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float)); yading@10: yading@10: for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) { yading@10: out[i] = 0.0; yading@10: for (j = 0; j <= HB_FIR_SIZE; j++) yading@10: out[i] += data[i + j] * fir_coef[j]; yading@10: } yading@10: yading@10: memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float)); yading@10: } yading@10: #endif /* hb_fir_filter */ yading@10: yading@10: /** yading@10: * Update context state before the next subframe. yading@10: */ yading@10: static void update_sub_state(AMRWBContext *ctx) yading@10: { yading@10: memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE], yading@10: (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float)); yading@10: yading@10: memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float)); yading@10: memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float)); yading@10: yading@10: memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE], yading@10: LP_ORDER * sizeof(float)); yading@10: memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE], yading@10: UPS_MEM_SIZE * sizeof(float)); yading@10: memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k], yading@10: LP_ORDER_16k * sizeof(float)); yading@10: } yading@10: yading@10: static int amrwb_decode_frame(AVCodecContext *avctx, void *data, yading@10: int *got_frame_ptr, AVPacket *avpkt) yading@10: { yading@10: AMRWBContext *ctx = avctx->priv_data; yading@10: AVFrame *frame = data; yading@10: AMRWBFrame *cf = &ctx->frame; yading@10: const uint8_t *buf = avpkt->data; yading@10: int buf_size = avpkt->size; yading@10: int expected_fr_size, header_size; yading@10: float *buf_out; yading@10: float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing yading@10: float fixed_gain_factor; // fixed gain correction factor (gamma) yading@10: float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use yading@10: float synth_fixed_gain; // the fixed gain that synthesis should use yading@10: float voice_fac, stab_fac; // parameters used for gain smoothing yading@10: float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis yading@10: float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band yading@10: float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis yading@10: float hb_gain; yading@10: int sub, i, ret; yading@10: yading@10: /* get output buffer */ yading@10: frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k; yading@10: if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) yading@10: return ret; yading@10: buf_out = (float *)frame->data[0]; yading@10: yading@10: header_size = decode_mime_header(ctx, buf); yading@10: if (ctx->fr_cur_mode > MODE_SID) { yading@10: av_log(avctx, AV_LOG_ERROR, yading@10: "Invalid mode %d\n", ctx->fr_cur_mode); yading@10: return AVERROR_INVALIDDATA; yading@10: } yading@10: expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1; yading@10: yading@10: if (buf_size < expected_fr_size) { yading@10: av_log(avctx, AV_LOG_ERROR, yading@10: "Frame too small (%d bytes). Truncated file?\n", buf_size); yading@10: *got_frame_ptr = 0; yading@10: return AVERROR_INVALIDDATA; yading@10: } yading@10: yading@10: if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID) yading@10: av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n"); yading@10: yading@10: if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */ yading@10: avpriv_request_sample(avctx, "SID mode"); yading@10: return AVERROR_PATCHWELCOME; yading@10: } yading@10: yading@10: ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame), yading@10: buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]); yading@10: yading@10: /* Decode the quantized ISF vector */ yading@10: if (ctx->fr_cur_mode == MODE_6k60) { yading@10: decode_isf_indices_36b(cf->isp_id, ctx->isf_cur); yading@10: } else { yading@10: decode_isf_indices_46b(cf->isp_id, ctx->isf_cur); yading@10: } yading@10: yading@10: isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past); yading@10: ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1); yading@10: yading@10: stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final); yading@10: yading@10: ctx->isf_cur[LP_ORDER - 1] *= 2.0; yading@10: ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER); yading@10: yading@10: /* Generate a ISP vector for each subframe */ yading@10: if (ctx->first_frame) { yading@10: ctx->first_frame = 0; yading@10: memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double)); yading@10: } yading@10: interpolate_isp(ctx->isp, ctx->isp_sub4_past); yading@10: yading@10: for (sub = 0; sub < 4; sub++) yading@10: ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER); yading@10: yading@10: for (sub = 0; sub < 4; sub++) { yading@10: const AMRWBSubFrame *cur_subframe = &cf->subframe[sub]; yading@10: float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k; yading@10: yading@10: /* Decode adaptive codebook (pitch vector) */ yading@10: decode_pitch_vector(ctx, cur_subframe, sub); yading@10: /* Decode innovative codebook (fixed vector) */ yading@10: decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih, yading@10: cur_subframe->pul_il, ctx->fr_cur_mode); yading@10: yading@10: pitch_sharpening(ctx, ctx->fixed_vector); yading@10: yading@10: decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode, yading@10: &fixed_gain_factor, &ctx->pitch_gain[0]); yading@10: yading@10: ctx->fixed_gain[0] = yading@10: ff_amr_set_fixed_gain(fixed_gain_factor, yading@10: ctx->celpm_ctx.dot_productf(ctx->fixed_vector, yading@10: ctx->fixed_vector, yading@10: AMRWB_SFR_SIZE) / yading@10: AMRWB_SFR_SIZE, yading@10: ctx->prediction_error, yading@10: ENERGY_MEAN, energy_pred_fac); yading@10: yading@10: /* Calculate voice factor and store tilt for next subframe */ yading@10: voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0], yading@10: ctx->fixed_vector, ctx->fixed_gain[0], yading@10: &ctx->celpm_ctx); yading@10: ctx->tilt_coef = voice_fac * 0.25 + 0.25; yading@10: yading@10: /* Construct current excitation */ yading@10: for (i = 0; i < AMRWB_SFR_SIZE; i++) { yading@10: ctx->excitation[i] *= ctx->pitch_gain[0]; yading@10: ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i]; yading@10: ctx->excitation[i] = truncf(ctx->excitation[i]); yading@10: } yading@10: yading@10: /* Post-processing of excitation elements */ yading@10: synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain, yading@10: voice_fac, stab_fac); yading@10: yading@10: synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector, yading@10: spare_vector); yading@10: yading@10: pitch_enhancer(synth_fixed_vector, voice_fac); yading@10: yading@10: synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain, yading@10: synth_fixed_vector, &ctx->samples_az[LP_ORDER]); yading@10: yading@10: /* Synthesis speech post-processing */ yading@10: de_emphasis(&ctx->samples_up[UPS_MEM_SIZE], yading@10: &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem); yading@10: yading@10: ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE], yading@10: &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles, yading@10: hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE); yading@10: yading@10: upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE], yading@10: AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx); yading@10: yading@10: /* High frequency band (6.4 - 7.0 kHz) generation part */ yading@10: ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples, yading@10: &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles, yading@10: hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE); yading@10: yading@10: hb_gain = find_hb_gain(ctx, hb_samples, yading@10: cur_subframe->hb_gain, cf->vad); yading@10: yading@10: scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain); yading@10: yading@10: hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k], yading@10: hb_exc, ctx->isf_cur, ctx->isf_past_final); yading@10: yading@10: /* High-band post-processing filters */ yading@10: hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem, yading@10: &ctx->samples_hb[LP_ORDER_16k]); yading@10: yading@10: if (ctx->fr_cur_mode == MODE_23k85) yading@10: hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem, yading@10: hb_samples); yading@10: yading@10: /* Add the low and high frequency bands */ yading@10: for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) yading@10: sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15)); yading@10: yading@10: /* Update buffers and history */ yading@10: update_sub_state(ctx); yading@10: } yading@10: yading@10: /* update state for next frame */ yading@10: memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0])); yading@10: memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float)); yading@10: yading@10: *got_frame_ptr = 1; yading@10: yading@10: return expected_fr_size; yading@10: } yading@10: yading@10: AVCodec ff_amrwb_decoder = { yading@10: .name = "amrwb", yading@10: .type = AVMEDIA_TYPE_AUDIO, yading@10: .id = AV_CODEC_ID_AMR_WB, yading@10: .priv_data_size = sizeof(AMRWBContext), yading@10: .init = amrwb_decode_init, yading@10: .decode = amrwb_decode_frame, yading@10: .capabilities = CODEC_CAP_DR1, yading@10: .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"), yading@10: .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT, yading@10: AV_SAMPLE_FMT_NONE }, yading@10: };